fix(hir): class static-field reflection remnant

[proggeramlug]

Summary

Closes the static-field leaf of the test262 language/{statements,expressions}/class remnant by fixing three shared root causes behind the class/elements/static-* cluster. All three converge on a single observation: a class constructor is an INT32-tagged ClassRef, not a heap object, so the generic object reflection paths skipped it.

test262 language/statements/class + language/expressions/class:
pass 5088 → 5094 (+6), runtime-fail 196 → 190, diff/compile-fail unchanged → zero regressions.

Root causes & fixes

1. Object.keys(C) / for (k in C) returned [].
   js_object_keys_value had no ClassRef arm; the INT32 value isn't a pointer, so it fell through to an empty array. Added a class-ref arm returning the enumerable static-field keys from CLASS_DYNAMIC_PROPS in ECMA order (#-private and the non-enumerable length/name/prototype + static methods excluded), via a new class_own_enumerable_field_names helper. Both Object.keys and the for-in desugar funnel through this function.

2. Object.prototype.propertyIsEnumerable.call(C, 'f') returned false.
   extract_obj_ptr nulls out on the INT32 ClassRef payload, so every key was reported non-enumerable. Added a class-ref string-key arm: a registered static field → enumerable, everything else → non-enumerable. This is the third leg of verifyProperty's isEnumerable (for-in + hasOwnProperty + propertyIsEnumerable); only with Support custom menu bar items #1 and linux compilation and README #2 does the descriptor cluster pass.

3. Uninitialized non-computed static fields (static foo;, static 0;) registered too late.
   They were handled by init_static_fields_late, which runs after user statements — so Object.keys(C) / getOwnPropertyDescriptor(C, "foo") immediately after the class declaration saw nothing. Their value is the compile-time constant undefined, and a class name is in TDZ before its declaration, so registration moved to init_static_fields_early (before user code). The late else-branch was removed — it also clobbered any C.foo = … the program performed before module-init end.

Out of scope (remaining static-field tail)

• Computed-key static fields with a string key (static ['constructor'];) route through the symbol side-table rather than CLASS_DYNAMIC_PROPS (string-vs-symbol key routing).
• Numeric-key static-field reads (static 0 = 'bar'; C[0]).
• Computed-key double-evaluation (static [i++] = i++).

Validation

• Class subset: see counts above (zero regressions — pass strictly up, no bucket grew).
• native-region-proof gate: unchanged vs origin/main. Its sole failure (h1_buffer_alias_negative, an unrelated HTTP buffer-alias codegen proof) reproduces identically on a clean origin/main build — pre-existing, not introduced here.
• cargo fmt --all clean; file-size gate green (all edited large files already allowlisted).

Files

• crates/perry-runtime/src/object/field_get_set.rs — js_object_keys_value class-ref arm
• crates/perry-runtime/src/object/object_ops.rs — propertyIsEnumerable class-ref arm
• crates/perry-runtime/src/object/class_registry.rs — class_own_enumerable_field_names
• crates/perry-runtime/src/object/descriptors.rs — sort_property_names_ecma made pub(crate)
• crates/perry-codegen/src/codegen/helpers.rs — early registration of no-init static fields